Back in 2022 and 2023 there were two big branches of machine learning happening at Meta1. The LLM work that led to Llama was a clean, smooth stack of repeated Transformer modules; the recommendatio…